scripts/build.py

#!/usr/bin/env python
r"""
Builds a WebDataset from the Cityscapes Video dataset.

Adapted from the `WebDataset documentation<https://github.com/webdataset/webdataset/>`_.
"""

import itertools
import collections
import typing as T
from pprint import pformat
import argparse
import multiprocessing as mp
import tarfile
import pandas as pd
from io import BytesIO
import json

from pathlib import Path
from tqdm import tqdm


def parse_args():
    ap = argparse.ArgumentParser(
        description="Build a WebDataset from the Cityscapes Video dataset."
    )

    # Flags and optional
    ap.add_argument(
        "--shard-size",
        "-s",
        type=int,
        default=10,
        help=("Number of sequences per shard."),
    )
    ap.add_argument(
        "--name",
        "-n",
        type=str,
        default="csvps",
        help=(
            "Name of the dataset. This will be used as the prefix for the tar files."
        ),
    )
    ap.add_argument(
        "--variant",
        type=str,
        default="",
        help=(
            "When passing different manifest variants, this will be used to postfix "
            "each split such that the resulting dataset name is unique."
        ),
    )
    ap.add_argument(
        "--force", "-f", action="store_true", help="Overwrite existing data."
    )
    ap.add_argument(
        "--splits", nargs="+", default=["train", "val", "test"], help="Splits to build."
    )
    ap.add_argument("--compression", "-c", default="", help="Compression to use")

    # Positional
    ap.add_argument("manifest", type=Path, help="Path to the manifest CSV file.")
    ap.add_argument("data", type=Path, help="Path to the Cityscapes Video dataset.")
    ap.add_argument("output", type=Path, help="Path to the output directory.")

    rt = ap.parse_args()

    # Validation
    if rt.shard_size < 1:
        ap.error("Shard size must be a positive integer.")
    if rt.name == "":
        ap.error("Name must be a non-empty string.")
    if not rt.name.isalnum() and not rt.name.islower():
        ap.error("Name must be a lowercase alpha-numeric string.")
    if rt.variant != "" and not rt.variant.isalnum() and not rt.variant.islower():
        ap.error("Variant must be a lowercase alpha-numeric string.")
    if not rt.manifest.exists():
        ap.error(f"Manifest file not found: {rt.manifest}")
    if not rt.data.exists():
        ap.error(f"Data directory not found: {rt.data}")
    if not rt.output.exists():
        rt.output.mkdir(parents=True)
        print(f"Created output directory: {rt.output}")

    return rt


PAD_TO: T.Final[int] = 6  # 06-padding is given by the dataset and should not be changed


def pad_number(n: int) -> str:
    r"""
    For sorting, numbers are padded with zeros to a fixed width.
    """
    if not isinstance(n, int):
        msg = f"Expected an integer, got {n} of type {type(n)}"
        raise TypeError(msg)
    return f"{n:0{PAD_TO}d}"


def read_timestamp(path: Path) -> int:
    with path.open("r") as f:
        ts = f.read().strip()
    if not ts.isdigit():
        msg = f"Expected a timestamp, got {ts} from {path}"
        raise ValueError(msg)
    return int(ts)


def write_bytes(tar: tarfile.TarFile, bt: bytes, arc: str):
    r""" "
    Simple utility to write the bytes (e.g. metadata json) directly from memory to
    the tarfile, since these do not exist as a file.
    """
    with BytesIO() as buf:
        buf.write(bt)

        # The TarInfo object must be created manually since the meta-data
        # JSON is written to a buffer (BytesIO) and not a file.
        tar_info = tarfile.TarInfo(arc)
        tar_info.size = buf.tell()  # number of bytes written

        # Reset the buffer to the beginning before adding it to the tarfile
        buf.seek(0)

        tar.addfile(tar_info, buf)


def find_sequence_files(
    seq: int,
    group: pd.DataFrame,
    *,
    data_dir: Path,
    dataset_name: str,
    compression: str,
    missing_ok: bool = False,
    frame_inputs: T.Sequence[str] = ("image.png", "vehicle.json"),
    frame_annotations: T.Sequence[str] = ("panoptic.png", "depth.tiff"),
    sequence_data: T.Sequence[str] = ("camera.json",),
    separator: str = "/",
) -> T.Iterator[tuple[Path | bytes, str]]:
    seq_pad = pad_number(seq)
    seq_dir = data_dir / seq_pad

    group = group.sort_values("frame")

    # Add frame-wise data
    primary_keys = group.index.tolist()
    frame_numbers = list(map(pad_number, group["frame"].tolist()))

    for i, meta in enumerate(
        group.drop(columns=["sequence", "frame", "split"]).to_dict(
            orient="records", index=True
        )
    ):
        frame_06 = frame_numbers[i]
        is_ann = meta["is_annotated"]

        # Write primary key
        meta["primary_key"] = primary_keys[i]

        # Add files to the tarfile
        for var in frame_inputs + frame_annotations:
            path_file = seq_dir / f"{frame_06}.{var}"
            if not path_file.exists():
                if missing_ok or (var in frame_annotations and not is_ann):
                    continue  # missing annotation OK
                msg = f"File not found: {path_file}"
                raise FileNotFoundError(msg)

            yield (
                path_file,
                separator.join(
                    (
                        dataset_name,
                        # {seq}.{frame}.{var}.{ext}
                        path_file.relative_to(data_dir).as_posix().replace("/", "."),
                    )
                ),
            )

        # Add the timestamp to the meta-data if it exists
        path_ts = seq_dir / f"{frame_06}.timestamp.txt"
        if not path_ts.exists():
            if not missing_ok:
                msg = f"Timestamp file not found: {path_ts}"
                raise FileNotFoundError(msg)
            meta["timestamp"] = None
        else:
            meta["timestamp"] = read_timestamp(path_ts)

        # Write frame metadata
        yield (
            json.dumps(meta).encode("utf-8"),
            f"{dataset_name}/{seq_pad}.{frame_06}.metadata.json",
        )

    # Add sequence-wise files {seq}.{var}.{ext}, e.g. 000000.camera.json
    for var in sequence_data:
        path_file = seq_dir.with_suffix("." + var)
        if not path_file.exists():
            if missing_ok:
                continue
            msg = f"File not found: {path_file}"
            raise FileNotFoundError(msg)

        yield (
            path_file,
            separator.join(
                (
                    dataset_name,
                    # {seq}.{var}.{ext}
                    path_file.relative_to(data_dir).as_posix(),
                )
            ),
        )

    # Write frames array
    yield (
        json.dumps(frame_numbers).encode("utf-8"),
        f"{dataset_name}/{seq_pad}.frames.json",
    )


def run_collector(
    seq: int, group: pd.DataFrame, kwargs: dict
) -> tuple[int, list[tuple[Path | bytes, str]]]:
    r"""
    Worker that collects the files for a single sequence.
    """
    return (seq, list(find_sequence_files(seq, group, **kwargs)))


def run_writer(
    tar_path: Path, items: list[list[tuple[Path | bytes, str]]], compression: str = "gz"
) -> None:
    r"""
    Worker that writes the files to a tar archive.
    """
    if compression != "":
        tar_path = tar_path.with_suffix(f".tar.{compression}")
        write_mode = f"w:{compression}"
    else:
        tar_path = tar_path.with_suffix(".tar")
        write_mode = "w"

    with tarfile.open(tar_path, write_mode) as tar:
        for item in itertools.chain.from_iterable(items):
            try:
                path, arc = item
            except ValueError:
                msg = f"Expected a tuple of length 2, got {item}"
                raise ValueError(msg)

            if isinstance(path, Path):
                tar.add(path, arcname=arc)
            else:
                write_bytes(tar, path, arc)


def build_shard(
    mfst: pd.DataFrame,
    *,
    tar_dir: Path,
    shard_size: int,
    **kwargs,
):
    # Make dirs
    tar_dir.mkdir(exist_ok=True, parents=True)

    write_log = collections.defaultdict(list)

    # Create a list of all sequences
    # groups = [(seq, group) for seq, group in mfst.groupby("sequence")]
    # shards = [groups[i : i + shard_size] for i in range(0, len(groups), shard_size)
    n_groups = len(mfst["sequence"].unique())
    n_shards = n_groups // shard_size

    targets = [None] * n_groups

    # Start a multiprocessing pool
    n_proc = min(mp.cpu_count(), 16)
    with mp.Pool(n_proc) as pool:
        write_jobs: list[mp.AsyncResult] = []

        # Data collection
        with tqdm(total=n_groups, desc="Collecting data") as pbar_group:
            for seq, files in pool.starmap(
                run_collector,
                [(seq, group, kwargs) for seq, group in mfst.groupby("sequence")],
                chunksize=min(8, shard_size),
            ):
                assert targets[seq] is None, f"Duplicate sequence: {seq}"

                pbar_group.update()

                # Write to the file specs list
                targets[seq] = files

                # Get a view of only the current shards's files
                shard_index = seq // shard_size
                shard_offset = shard_index * shard_size
                shard_specs = targets[shard_offset : shard_offset + shard_size]

                # Pad the shard index
                shard_06 = pad_number(shard_index)

                write_log[shard_06].append(pad_number(seq))

                # If the shard is fully populated, write it to a tar file in another process
                if all(s is not None for s in shard_specs):
                    tar_path = tar_dir / shard_06

                    write_jobs.append(
                        pool.apply_async(
                            run_writer,
                            (tar_path, shard_specs, ""),
                        )
                    )

        # Wait for write-workers to finish generating the TAR files
        with tqdm(total=n_shards, desc="Writing shards") as pbar_shard:
            for j in write_jobs:
                j.get()
                pbar_shard.update()

        pool.close()
        pool.join()

    print("Created shard files:\n" + pformat(dict(write_log)))


def main():
    args = parse_args()
    manifest = pd.read_csv(args.manifest, index_col="primary_key")

    # For each split, build a tar archive containing the sorted files
    for split in args.splits:
        split_out = "-".join([s for s in (split, args.variant) if len(s) > 0])
        tar_dir = args.output / split_out

        if tar_dir.exists():
            if args.force:
                print(f"Removing existing dataset: {tar_dir}")
                for f in tar_dir.glob("*.tar"):
                    f.unlink()
            else:
                msg = f"Dataset already exists: {tar_dir}"
                raise FileExistsError(msg)

        print(f"Generating {split_out} split...")

        build_shard(
            manifest[manifest["split"] == split],
            tar_dir=tar_dir,
            data_dir=args.data / split,
            shard_size=args.shard_size,
            dataset_name=f"{args.name}-{split_out}",
            missing_ok=True,
            compression=args.compression
        )


if __name__ == "__main__":
    main()